WingCore: Pipeline Convolutional Neural Network Accelerator for Object Detection Using the Winograd Algorithm

Denny H. T. Nugroho; Trio Adiono; Infall Syafalni; Nana Sutisna

doi:10.1109/ISESD56103.2022.9980770

← Back to search

WingCore: Pipeline Convolutional Neural Network Accelerator for Object Detection Using the Winograd Algorithm

Closed

Denny H. T. Nugroho, Trio Adiono, Infall Syafalni, Nana Sutisna

2022 ISESD 2022 - 2022 International Symposium on Electronics and Smart Devices, Proceeding Conference paper Cited by 1

Abstract

Convolutional Neural Network (CNN) has been widely used in computer vision problems in recent years. Because of their excellent performance, energy economy, and versatility, FPGAs have been intensively investigated for use in accelerating CNNs. In this paper, we present WingCore, which uses the fast Winograd method to improve FPGA-based neural network accelerators. We use the weight stationary method and the addition process after element-wise multiplication to improve accelerator performance. We evaluate the hardware architecture on PYNQ Z2 platform. When compared to the CPU, our design achieves a 216.16 × speedup on Yolov3-Tiny layer 1. © 2022 IEEE.

Affiliations

School of Electrical Engineering and Informatics, Indonesia; Sumatera Institute of Technology, Electrical Engineering Department, Indonesia; University Center of Excellence on Microelectronics, Bandung Institute of Technology, Indonesia

View full text (DOI) →

Research at a Glance

Research at a Glance

Topics & SDG Alignment

Collaboration

Author Profile (Selected)

References Overview

Journal & Source

Metadata & Integrity